A GAN noise modeling based blind denoising method for guided waves. (January 2022)

• Record Type: Journal Article
• Authors: Cui, Xiushi; Li, Dongsheng; Li, Ziqi; Ou, Jinping
• Is Part Of: Measurement. Volume 188(2022)
• Publication Date: 2022-01
• Subjects: Guided waves -- Generative adversarial network -- Denoising autoencoder -- Prior information -- non-Gaussian features; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.110596
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: The combination of GAN and AE for signal denoising is a novel idea. Double sliding window estimate method is used to extract the noise block in the signal. LSGAN is used to learn the distribution features of noise and generate training samples. The SNR and the degree of restoration of signal features are used to evaluate the denoising effect. Abstract: In the detection using guided waves, the signal often carries a high level of non-Gaussian noise. The traditional denoising method cannot estimate and use the prior information of the signal, which leads to poor denoising effect. To tackle this problem, this paper proposed a denoising network based on the combination of generative adversarial network (GAN) and autoencoder (AE). First, GAN is used to estimate the distribution characteristics of the extracted noise and generate samples. Second, according to the characteristics of the guided wave, a pair of datasets are generated to train DAE network. The trained denoising AE has strong robustness. As a result, the proposed GAN-AE based denoiser (GAD) can effectively can effectively reduce the noise level and has the ability to accurately recover the peak time of the wave packet. In particular, the proposed method significantly outperforms conventional denoising methods in low signal-to-noise (SNR) conditions.